知识卡片

普通读书笔记卡 · 8707 · 生成模型:GPT-6

首轮召回应覆盖所有必要证据

理解等级:中级

初始检索以合理成本取回可能有用候选,后续重排只能在该集合中选择。对多跳和多条件任务,命中一篇相关文档未必足够,需评估所需证据集合是否齐备。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      of the RAG pipeline. It is easy to see why: all RAG applications are bottlenecked by the quality of retrieval. Even if you are working with the world’s best language model, you won’t be able to get the correct results if the retrieval step didn’t retrieve the correct documents needed to answer the query. Therefore, this step of the pipeline should focus on increasing recall.

    参考来源

    - 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html) - 结论依据:摘录中关于首轮召回应覆盖所有必要证据的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:of the RAG pipeline. It is easy to see why: all RAG applications are bottlenecked by the quality of retrieval. Even if you are working with the world’s best language model, you won’t be able to get the correct results if the retrieval step didn’t retrieve the correct documents needed to answer the query. Therefore, this step of the pipeline should focus on increasing recall.